
Scheme	`FP8_DYNAMIC` — weights static per-channel FP8, activations dynamic per-token FP8
Calibration	none needed; dynamic activation scales are computed at inference time
Left in bf16	`lm_head`, and the token embeddings
Weights before	7.49 GiB
Weights after	4.11 GiB (45% smaller)

Embeddings are the pooled last hidden state; the head is unused and left unquantised. Cosine similarity is scale-sensitive, so compare rankings rather than raw distances when validating this checkpoint.

Quality gate

WikipediaRetrievalMultilingual from MTEB — retrieval over Wikipedia in 16 languages, scored by ndcg_at_10.

language	bf16	FP8	Δ
de	0.9066	0.9065	-0.0001
en	0.9344	0.9353	+0.0009
it	0.8993	0.8994	+0.0002
mean	0.9134	0.9138	+0.0003

Why

Serving Qwen/Qwen3-Embedding-4B in bf16 leaves little room for KV cache on a small GPU: the weights take what the cache needs, and the context length has to be cut until it fits. Halving the weights gives that memory back — the same card serves a longer context without any other change.

Serving

bash

vllm serve DCC-BS/Qwen3-Embedding-4B-FP8-Dynamic \
  --task embed \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.85

The checkpoint is in compressed-tensors format, so vLLM detects the quantisation from config.json; no extra flag is required.